Deep learning approach for crop-weed segmentation in peanut cultivation using PSPEdgeWeedNet

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Weed management is a major agricultural challenge due to weed competition and visual similarity with crops.
• Automated weed detection systems aim to reduce costs and herbicide reliance but struggle with segmentation accuracy.
• Variations in lighting and field conditions further complicate precise crop-weed differentiation.

Purpose of the Study:

• To develop a novel deep learning architecture for accurate semantic segmentation of crops and weeds in peanut fields.
• To enhance boundary localization and delineation between crops and weeds using an edge-aware mechanism.
• To improve the robustness and accuracy of automated weed detection systems in complex agricultural settings.

Main Methods:

• Proposed PSPEdgeWeedNet, an edge-aware deep learning architecture for semantic segmentation.
• Introduced a dedicated edge detection branch to improve boundary delineation.
• Utilized Conditional Random Fields (CRFs) for post-processing enhancement of segmentation boundaries.
• Trained models on a peanut field dataset using class-weighted loss functions to handle class imbalance.

Main Results:

• PSPEdgeWeedNet significantly outperformed state-of-the-art models like PSPNet, SegNet, UNet, DeepLabv3, Swin-Unet, and ViT-based models.
• Achieved superior performance across key metrics including Intersection over Union (IoU), precision, recall, and F1-score.
• Demonstrated the effectiveness of edge-aware mechanisms in improving segmentation accuracy for weed detection.

Conclusions:

• Incorporating edge-aware mechanisms is critical for enhancing semantic segmentation in agricultural applications.
• PSPEdgeWeedNet offers a robust and accurate solution for automated weed detection in peanut cultivation.
• The study highlights the potential of advanced deep learning for precision agriculture and sustainable farming practices.